If you run more than one cluster, administration means logging into each one separately. With 9.1.5 Qontrol handles the whole fleet from a single console.
The MCP server is not part of that. It stays cluster local, it is off by default, and the tool set is configurable, so you decide exactly what an AI assistant may see and do. Privacy first.
The new piece is qontrol-agent, a headless Qontrol that runs next to a cluster and opens a secure connection outbound to a remote Qontrol server. No inbound ports are needed, so clusters behind a firewall, in another data center, or in a cloud region all work the same way. On the Qontrol server you create the keys, and you remove access again when a cluster should no longer connect.
The interface itself does not change. The only new element is a dropdown at the top where you select the cluster you want to work with. Everything below it stays exactly what you already know: add queues, edit calendars, set up resource quota sets, and everything else Qontrol does locally. All from one place.
Resource maps got a real upgrade. RSMAPs now carry characteristics, typed metadata attached to each device instance: device files, memory, socket affinity, bandwidth. The scheduler uses that information during placement, and the execution daemon hands the resulting device list to systemd as DeviceAllow, so the kernel enforces the boundary. This is not GPU only. NPUs and other accelerators, FPGAs, high speed network interfaces, and storage devices work the same way.
Most of this release you will never see. A lot of work went into the backend: communication and job submission under load, more reliable job tracking in the execution daemon, accounting fixes for long running and parallel environment jobs, thread safety, and support for newer platforms like Ubuntu 26.04 and gcc 15. Not much to show in a screenshot. That is where cluster stability comes from.
qtelemetry gained an execution host mode. Per job metrics, memory, I/O, and CPU over time, exported directly from the execution hosts. That means no extra query load on qmaster, no matter how many jobs you monitor. Per job monitoring is the interesting part here, and the dashboards are ready to use:
Gridware Cluster Scheduler Execution Hosts
There is more in this release than fits here, including the containerized Qontrol deployment with Dockerfile and helm chart. The full release post is worth reading: Open Cluster Scheduler and Gridware Cluster Scheduler 9.1.5